Darknet shape inference (#370) (#71)

src/darknet.js

@@ -70,8 +70,8 @@ darknet.Graph = class {
         let nu = 0;
         while (lines.length > 0) {
             nu++;
-            let line = lines.shift();
-            line = line.replace(/\s/g, '');
+            const text = lines.shift();
+            const line = text.replace(/\s/g, '');
             if (line.length > 0) {
                 switch (line[0]) {
                     case '#':
@@ -85,10 +85,13 @@ darknet.Graph = class {
                         break;
                     }
                     default: {
+                        if (!section || line[0] < 0x20 || line[0] > 0x7E) {
+                            throw new darknet.Error("Invalid cfg '" + text.replace(/[^\x20-\x7E]+/g, '').trimStart().trimEnd() + "' at line " + nu.toString() + ".");
+                        }
                         if (section) {
                             let property = line.split('=');
                             if (property.length != 2) {
-                                throw new darknet.Error("Invalid cfg '" + line + "' at line " + nu.toString() + ".");
+                                throw new darknet.Error("Invalid cfg '" + text.replace(/[^\x20-\x7E]+/g, '').trimStart().trimEnd() + "' at line " + nu.toString() + ".");
                             }
                             let key = property[0].trim();
                             let value = property[1].trim();
@@ -100,94 +103,353 @@ darknet.Graph = class {
             }
         }
 
-        for (let section of sections) {
-            section.values = {};
-            const schema = metadata.getSchema(section.type);
-            if (schema && schema.attributes) {
-                for (let attribute of schema.attributes) {
-                    if (attribute.name) {
-                        if (section.options[attribute.name] !== undefined) {
-                            switch (attribute.type) {
-                                case 'int32':
-                                    section.values[attribute.name] = parseInt(section.options[attribute.name], 10);
-                                    break;
-                                case 'float32':
-                                    section.values[attribute.name] = parseFloat(section.options[attribute.name]);
-                                    break;
-                                case 'string':
-                                    section.values[attribute.name] = section.options[attribute.name];
-                                    break;
-                            }
-                        }
-                        else if (attribute.default !== undefined) {
-                            section.values[attribute.name] = attribute.default
-                        }
-                    }
-                }
+        const option_find_int = (options, key, defaultValue) => {
+            const value = options[key];
+            return value !== undefined ? parseInt(value, 10) : defaultValue;
+        };
+
+        const option_find_str = (options, key, defaultValue) => {
+            const value = options[key];
+            return value !== undefined ? value : defaultValue;
+        };
+
+        let params = {};
+
+        const net = sections.shift();
+        switch (net.type) {
+            case 'net':
+            case 'network': {
+                params.h = option_find_int(net.options, 'height', 0);
+                params.w = option_find_int(net.options, 'width', 0);
+                params.c = option_find_int(net.options, 'channels', 0);
+                params.inputs = option_find_int(net.options, 'inputs', params.h * params.w * params.c);
+                break;
             }
         }
 
-        if (sections.length === 0) {
-            throw new darknet.Error('Config file has no sections.');
-        }
-
-        let net = sections.shift();
-        if (net.type !== 'net' && net.type !== 'network') {
-            throw new darknet.Error('First section must be [net] or [network].');
-        }
-
-        const inputType = new darknet.TensorType('float32', new darknet.TensorShape([ net.values.width, net.values.height, net.values.channels ]));
-
+        const inputType = params.w && params.h && params.c ?
+            new darknet.TensorType('float32', new darknet.TensorShape([ params.w, params.h, params.c ])) :
+            new darknet.TensorType('float32', new darknet.TensorShape([ params.inputs ]));
         const inputName = 'input';
-        this._inputs.push(new darknet.Parameter(inputName, true, [
-            new darknet.Argument(inputName, inputType, null)
-        ]));
+        params.arguments = [ new darknet.Argument(inputName, inputType, null) ];
+        this._inputs.push(new darknet.Parameter(inputName, true, params.arguments));
 
-        for (let i = 0; i < sections.length; i++) {
-            sections[i]._outputs = [ i.toString() ];
+        if (sections.length === 0) {
+            throw new darknet.Error('Config file has no sections.');
         }
 
-        let inputs = [ inputName ];
+        let infer = true;
         for (let i = 0; i < sections.length; i++) {
-            const layer = sections[i];
-            layer._inputs = inputs;
-            inputs = [ i.toString() ];
-            switch (layer.type) {
+            let section = sections[i];
+            section.layer = {};
+            section.tensors = [];
+            section.inputs = [];
+            section.outputs = [];
+            const options = section.options;
+            let layer = section.layer;
+            section.inputs = section.inputs.concat(params.arguments);
+            section.outputs.push(new darknet.Argument(i.toString(), null, null));
+            switch (section.type) {
                 case 'shortcut':
                 case 'sam':
                 case 'scale_channels': {
-                    let from = Number.parseInt(layer.options.from, 10);
-                    from = (from >= 0) ? from : (i + from);
-                    const shortcut = sections[from];
-                    if (shortcut) {
-                        layer._inputs.push(shortcut._outputs[0]);
+                    let index = option_find_int(options, 'from', 0);
+                    if (index < 0) {
+                        index = i + index;
+                    }
+                    const from = sections[index];
+                    if (from) {
+                        section.inputs.push(from.outputs[0]);
+                        section.from = from;
                     }
+                    delete options.from;
                     break;
                 }
                 case 'route': {
-                    layer._inputs = [];
-                    const routes = layer.options.layers.split(',').map((route) => Number.parseInt(route.trim(), 10));
+                    section.inputs = [];
+                    section.input_sections = [];
+                    const routes = options.layers.split(',').map((route) => Number.parseInt(route.trim(), 10));
                     for (let j = 0; j < routes.length; j++) {
                         const index = (routes[j] < 0) ? i + routes[j] : routes[j];
                         const route = sections[index];
                         if (route) {
-                            layer._inputs.push(route._outputs[0]);
+                            section.inputs.push(route.outputs[0]);
+                            section.input_sections.push(route);
                         }
                     }
+                    delete options.layers;
                     break;
                 }
             }
+            if (infer) {
+                switch (section.type) {
+                    case 'convolutional':
+                    case 'deconvolutional': {
+                        const w = params.w;
+                        const h = params.h;
+                        const c = params.c;
+                        const size = option_find_int(options, 'size', 1);
+                        const n = option_find_int(options, 'filters', 1);
+                        const pad = option_find_int(options, 'pad', 0);
+                        const padding = pad ? (size >> 1) : option_find_int(options, 'padding', 0);
+                        const stride = option_find_int(options, 'stride', 1);
+                        const groups = option_find_int(options, 'groups', 1);
+                        layer.out_w = Math.floor((w + 2 * padding - size) / stride) + 1;
+                        layer.out_h = Math.floor((h + 2 * padding - size) / stride) + 1;
+                        layer.out_c = n;
+                        layer.outputs = layer.out_h * layer.out_w * layer.out_c;
+                        section.tensors.push({ name: 'weights', shape: [ Math.floor(c / groups), n, size, size ]});
+                        section.tensors.push({ name: 'biases', shape: [ n ]});
+                        section.outputs[0].type = new darknet.TensorType('float32', new darknet.TensorShape([ layer.out_h, layer.out_w, layer.out_c ]));
+                        break;
+                    }
+                    case 'connected': {
+                        const outputs = option_find_int(options, 'output', 1);
+                        section.tensors.push({ name: 'weights', shape: [ params.inputs, outputs ] });
+                        section.tensors.push({ name: 'biases', shape: [ outputs ] });
+                        section.outputs[0].type = new darknet.TensorType('float32', new darknet.TensorShape([ outputs ]));
+                        layer.out_h = 1;
+                        layer.out_w = 1;
+                        layer.out_c = outputs;
+                        layer.outputs = outputs;
+                        break;
+                    }
+                    case 'local': {
+                        const shape = section.inputs[0].type.shape.dimensions;
+                        if (shape[0] !== params.w || shape[1] !== params.h || shape[2] !== params.c) {
+                            throw darknet.Error('Layer before avgpool layer must output image.');
+                        }
+                        const n = option_find_int(options, 'filters' , 1);
+                        const size = option_find_int(options, 'size', 1);
+                        const stride = option_find_int(options, 'stride', 1);
+                        const pad = option_find_int(options, 'pad', 0);
+                        layer.out_h = Math.floor((params.h - (pad ? 1 : size)) / stride) + 1;
+                        layer.out_w = Math.floor((params.w - (pad ? 1 : size)) / stride) + 1;
+                        layer.out_c = n;
+                        layer.outputs = layer.out_w * layer.out_h * layer.out_c;
+                        section.tensors.push({ name: 'weights', shape: [ params.c, n, size, size, layer.out_h * layer.out_w ]});
+                        section.tensors.push({ name: 'biases', shape: [ layer.out_w * layer.out_h * layer.out_c ]});
+                        section.outputs[0].type = new darknet.TensorType('float32', new darknet.TensorShape([ layer.out_w, layer.out_h, layer.out_c ]));
+                        break;
+                    }
+                    case 'maxpool': {
+                        const shape = section.inputs[0].type.shape.dimensions;
+                        if (shape[0] !== params.w || shape[1] !== params.h || shape[2] !== params.c) {
+                            throw darknet.Error('Layer before maxpool layer must output image.');
+                        }
+                        const stride = option_find_int(options, 'stride', 1);
+                        const size = option_find_int(options, 'size', stride);
+                        const padding = option_find_int(options, 'padding', size - 1);
+                        layer.out_w = Math.floor((params.w + padding - size) / stride) + 1;
+                        layer.out_h = Math.floor((params.h + padding - size) / stride) + 1;
+                        layer.out_c = params.c;
+                        layer.outputs = layer.out_w * layer.out_h * layer.out_c;
+                        section.outputs[0].type = new darknet.TensorType('float32', new darknet.TensorShape([ layer.out_w, layer.out_h, layer.out_c ]));
+                        break;
+                    }
+                    case 'avgpool': {
+                        const shape = section.inputs[0].type.shape.dimensions;
+                        if (shape[0] !== params.w || shape[1] !== params.h || shape[2] !== params.c) {
+                            throw darknet.Error('Layer before avgpool layer must output image.');
+                        }
+                        layer.out_w = 1;
+                        layer.out_h = 1;
+                        layer.out_c = params.c;
+                        layer.outputs = layer.out_c;
+                        section.outputs[0].type = new darknet.TensorType('float32', new darknet.TensorShape([ layer.out_w, layer.out_h, layer.out_c ]));
+                        break;
+                    }
+                    case 'gru':
+                    case 'rnn': 
+                    case 'lstm':{
+                        const output = option_find_int(options, "output", 1);
+                        layer.outputs = output;
+                        section.outputs[0].type = new darknet.TensorType('float32', new darknet.TensorShape([ output ]));
+                        break;
+                    }
+                    case 'softmax':
+                    case 'dropout': {
+                        layer.out_w = params.w;
+                        layer.out_h = params.h;
+                        layer.out_c = params.c;
+                        layer.outputs = params.inputs;
+                        section.outputs[0].type = new darknet.TensorType('float32', new darknet.TensorShape([ layer.outputs ]));
+                        break;
+                    }
+                    case 'upsample': {
+                        const stride = option_find_int(options, 'stride', 2);
+                        layer.out_w = params.w * stride;
+                        layer.out_h = params.h * stride;
+                        layer.out_c = params.c;
+                        layer.outputs = layer.out_w * layer.out_h * layer.out_c;
+                        section.outputs[0].type = new darknet.TensorType('float32', new darknet.TensorShape([ layer.out_w, layer.out_h, layer.out_c ]));
+                        break;
+                    }
+                    case 'crop': {
+                        const shape = section.inputs[0].type.shape.dimensions;
+                        if (shape[0] !== params.w || shape[1] !== params.h || shape[2] !== params.c) {
+                            throw darknet.Error('Layer before crop layer must output image.');
+                        }
+                        const crop_height = option_find_int(options, 'crop_height', 1);
+                        const crop_width = option_find_int(options, 'crop_width', 1);
+                        layer.out_w = crop_width;
+                        layer.out_h = crop_height;
+                        layer.out_c = params.c;
+                        layer.outputs = layer.out_w * layer.out_h * layer.out_c;
+                        section.outputs[0].type = new darknet.TensorType('float32', new darknet.TensorShape([ layer.out_w, layer.out_h, layer.out_c ]));
+                        break;
+                    }
+                    case 'yolo': {
+                        const w = params.w;
+                        const h = params.h;
+                        const classes = option_find_int(options, 'classes', 20);
+                        const n = option_find_int(options, 'num', 1);
+                        layer.out_h = h;
+                        layer.out_w = w;
+                        layer.out_c = n * (classes + 4 + 1);
+                        layer.outputs = layer.out_h * layer.out_w * layer.out_c;
+                        section.outputs[0].type = new darknet.TensorType('float32', new darknet.TensorShape([ layer.out_w, layer.out_h, layer.out_c ]));
+                        break;
+                    }
+                    case 'region': {
+                        const coords = option_find_int(options, 'coords', 4);
+                        const classes = option_find_int(options, 'classes', 20);
+                        const num = option_find_int(options, 'num', 1);
+                        layer.outputs = params.h * params.w * num * (classes + coords + 1);
+                        section.outputs[0].type = new darknet.TensorType('float32', new darknet.TensorShape([ params.h, params.w, num, (classes + coords + 1) ]));
+                        break;
+                    }
+                    case 'cost': {
+                        layer.outputs = params.inputs;
+                        section.outputs[0].type = new darknet.TensorType('float32', new darknet.TensorShape([ layer.outputs ]));
+                        break;
+                    }
+                    case 'reorg': {
+                        const stride = option_find_int(options, 'stride', 1);
+                        const reverse = option_find_int(options, 'reverse', 0);
+                        const extra = option_find_int(options, 'extra', 0);
+                        if (reverse) {
+                            layer.out_w = params.w * stride;
+                            layer.out_h = params.h * stride;
+                            layer.out_c = Math.floor(params.c / (stride * stride));
+                        } 
+                        else {
+                            layer.out_w = Math.floor(params.w / stride);
+                            layer.out_h = Math.floor(params.h / stride);
+                            layer.out_c = params.c * (stride * stride);
+                        }
+                        layer.outputs = layer.out_h * layer.out_w * layer.out_c;
+                        if (extra) {
+                            layer.out_w = 0;
+                            layer.out_h = 0; 
+                            layer.out_c = 0;
+                            layer.outputs = (params.h * params.w * params.c) + extra;
+                        }
+                        section.outputs[0].type = new darknet.TensorType('float32', new darknet.TensorShape([ layer.outputs ]));
+                        break;
+                    }
+                    case 'scale_channels': {
+                        infer = false;
+                        break;
+                    }
+                    case 'route': {
+                        let layers = section.input_sections.map((section) => section.layer);
+                        layer.outputs = 0;
+                        for (let input_layer of layers) {
+                            layer.outputs += input_layer.outputs;
+                        }
+                        const first = layers.shift();
+                        layer.out_w = first.out_w;
+                        layer.out_h = first.out_h;
+                        layer.out_c = first.out_c;
+                        while (layers.length > 0) {
+                            const next = layers.shift();
+                            if (next.out_w === first.out_w && next.out_h === first.out_h) {
+                                layer.out_c += next.out_c;
+                            }
+                            else {
+                                layer.out_h = 0;
+                                layer.out_w = 0;
+                                layer.out_c = 0;
+                            }
+                        }
+                        section.outputs[0].type = new darknet.TensorType('float32', new darknet.TensorShape([ layer.out_h, layer.out_w, layer.out_c ]));
+                        break;
+                    }
+                    case 'shortcut': {
+                        const from = section.from;
+                        layer.w = from.layer.out_w;
+                        layer.h = from.layer.out_h;
+                        layer.c = from.layer.out_c;
+                        layer.out_w = params.w;
+                        layer.out_h = params.h;
+                        layer.out_c = params.c;
+                        layer.outputs = params.w * params.h * params.c;
+                        section.outputs[0].type = new darknet.TensorType('float32', new darknet.TensorShape([ params.w, params.h, params.c ]));
+                        break;
+                    }
+                    default: {
+                        infer = false;
+                        break;
+                    }
+                }
+                params.h = layer.out_h;
+                params.w = layer.out_w;
+                params.c = layer.out_c;
+                params.inputs = layer.outputs;
+            }
+            params.arguments = section.outputs;
+
+            const batch_normalize = option_find_int(section.options, 'batch_normalize', 0);
+            if (batch_normalize) {
+                let size = -1;
+                switch (section.type) {
+                    case 'convolutional': {
+                        size = option_find_int(options, 'filters', 1);
+                        break;
+                    }
+                    case 'crnn':
+                    case 'gru':
+                    case 'rnn':
+                    case 'lstm':
+                    case 'connected': {
+                        size = option_find_int(options, 'output', 1);
+                        break;
+                    }
+                }
+                if (size < 0) {
+                    throw new darknet.Error("Invalid batch_normalize size for '" + section.type + "'.");
+                }
+                let chain = {};
+                chain.type = 'batch_normalize';
+                chain.tensors = [
+                    { name: 'scale', shape: [ size ] },
+                    { name: 'mean', shape: [ size ] },
+                    { name: 'variance', shape: [ size ] }
+                ];
+                section.chain = section.chain || [];
+                section.chain.push(chain);
+            }
+    
+            const defaultActivation = section.type === 'shortcut' ? 'linear' : 'logistic';
+            const activation = option_find_str(section.options, 'activation', defaultActivation);
+            if (activation !== defaultActivation) {
+                let chain = {};
+                chain.type = activation;
+                section.chain = section.chain || [];
+                section.chain.push(chain);
+            }
         }
+
         for (let i = 0; i < sections.length; i++) {
             this._nodes.push(new darknet.Node(metadata, net, sections[i], i.toString()));
         }
 
         if (sections.length > 0) {
-            const lastLayer = sections[sections.length - 1];
-            for (let i = 0; i < lastLayer._outputs.length; i++) {
-                this._outputs.push(new darknet.Parameter('output' + (i > 1 ? i.toString() : ''), true, [
-                    new darknet.Argument(lastLayer._outputs[i], null, null)
-                ]));
+            const last = sections[sections.length - 1];
+            for (let i = 0; i < last.outputs.length; i++) {
+                const outputName = 'output' + (i > 1 ? i.toString() : '');
+                this._outputs.push(new darknet.Parameter(outputName, true, [ last.outputs[i] ]));
             }
         }
     }
@@ -245,6 +507,13 @@ darknet.Argument = class {
         return this._type;
     }
 
+    set type(value) {
+        if (this._type) {
+            throw new darknet.Error('Invalid argument type set operation.');
+        }
+        this._type = value;
+    }
+
     get initializer() {
         return this._initializer;
     }
@@ -252,72 +521,37 @@ darknet.Argument = class {
 
 darknet.Node = class {
 
-    constructor(metadata, net, layer, name) {
+    constructor(metadata, net, section, name) {
         this._name = name;
         this._metadata = metadata;
-        this._operator = layer.type;
+        this._operator = section.type;
         this._attributes = [];
         this._inputs = [];
         this._outputs = [];
         this._chain = [];
-        if (layer._inputs && layer._inputs.length > 0) {
-            this._inputs.push(new darknet.Parameter(layer._inputs.length <= 1 ? 'input' : 'inputs', true, layer._inputs.map((input) => {
-                return new darknet.Argument(input, null, null);
-            })));
+        if (section.inputs && section.inputs.length > 0) {
+            this._inputs.push(new darknet.Parameter(section.inputs.length <= 1 ? 'input' : 'inputs', true, section.inputs));
         }
-        if (layer._outputs && layer._outputs.length > 0) {
-            this._outputs.push(new darknet.Parameter(layer._outputs.length <= 1 ? 'output' : 'outputs', true, layer._outputs.map((output) => {
-                return new darknet.Argument(output, null, null);
-            })));
+        if (section.tensors && section.tensors.length > 0) {
+            for (let tensor of section.tensors) {
+                const type = new darknet.TensorType('float', new darknet.TensorShape(tensor.shape));
+                this._inputs.push(new darknet.Parameter(tensor.name, true, [
+                    new darknet.Argument('', null, new darknet.Tensor('', type) )
+                ]))
+            }
         }
-        switch (layer.type) {
-            case 'convolutional':
-            case 'deconvolutional':
-                this._initializer('biases', [ layer.values.filters ]);
-                this._initializer('weights', [ net.values.channels, layer.values.size, layer.values.size, layer.values.filters ]);
-                this._batch_normalize(metadata, net, layer, layer.values.filters);
-                this._activation(metadata, net, layer, 'logistic');
-                break;
-            case 'connected':
-                this._initializer('biases', [ layer.values.output ]);
-                this._initializer('weights');
-                this._batch_normalize(metadata, net, layer, layer.values.output);
-                this._activation(metadata, net, layer);
-                break;
-            case 'crnn':
-                this._batch_normalize(metadata, net, layer);
-                this._activation(metadata, net, layer);
-                break;
-            case 'rnn':
-                this._batch_normalize(metadata, net, layer, layer.values.output);
-                this._activation(metadata, net, layer);
-                break;
-            case 'gru':
-                this._batch_normalize(metadata, net, layer);
-                break;
-            case 'lstm':
-                this._batch_normalize(metadata, net, layer);
-                break;
-            case 'shortcut':
-                this._activation(metadata, net, layer);
-                break;
-            case 'batch_normalize':
-                this._initializer('scale', [ layer.values.size ]);
-                this._initializer('mean', [ layer.values.size ]);
-                this._initializer('variance', [ layer.values.size ]);
-                break;
+        if (section.outputs && section.outputs.length > 0) {
+            this._outputs.push(new darknet.Parameter(section.outputs.length <= 1 ? 'output' : 'outputs', true, section.outputs));
         }
-
-        switch (layer.type) {
-            case 'shortcut':
-                delete layer.options.from;
-                break;
-            case 'route':
-                delete layer.options.layers;
-                break;
+        if (section.chain) {
+            for (let chain of section.chain) {
+                this._chain.push(new darknet.Node(metadata, net, chain, ''));
+            }
         }
-        for (let key of Object.keys(layer.options)) {
-            this._attributes.push(new darknet.Attribute(metadata, this._operator, key, layer.options[key]));
+        if (section.options) {
+            for (let key of Object.keys(section.options)) {
+                this._attributes.push(new darknet.Attribute(metadata, this._operator, key, section.options[key]));
+            }
         }
     }
 
@@ -390,31 +624,6 @@ darknet.Node = class {
     get chain() {
         return this._chain;
     }
-
-    _initializer(name, shape) {
-        const id = this._name.toString() + '_' + name;
-        this._inputs.push(new darknet.Parameter(name, true, [
-            new darknet.Argument(id, null, new darknet.Tensor(id, shape))
-        ]));
-    }
-
-    _batch_normalize(metadata, net, layer, size) {
-        if (layer.values.batch_normalize === 1) {
-            const batch_normalize_layer = { type: 'batch_normalize', options: {}, values: { size: size || 0 }, _inputs: [], _outputs: [] };
-            this._chain.push(new darknet.Node(metadata, net, batch_normalize_layer, ''));
-        }
-        delete layer.options.batch_normalize;
-    }
-
-    _activation(metadata, net, layer) {
-        const attributeSchema = metadata.getAttributeSchema(layer.type, 'activation');
-        if (attributeSchema) {
-            if (layer.options.activation !== attributeSchema.default) {
-                this._chain.push(new darknet.Node(metadata, net, { type: layer.options.activation, options: {}, values: {}, _inputs: [], _outputs: [] }, ''));
-            }
-            delete layer.options.activation;
-        }
-    }
 };
 
 darknet.Attribute = class {
@@ -465,10 +674,13 @@ darknet.Attribute = class {
 
 darknet.Tensor = class {
 
-    constructor(id, shape) {
-        shape = shape || null;
+    constructor(id, type) {
         this._id = id;
-        this._type = new darknet.TensorType('?', new darknet.TensorShape(shape));
+        this._type = type;
+    }
+
+    get kind() {
+        return 'Tensor';
     }
 
     get name() {
@@ -515,6 +727,9 @@ darknet.TensorType = class {
 darknet.TensorShape = class {
 
     constructor(dimensions) {
+        if (dimensions.some((dimension) => !dimension)) {
+            throw new darknet.Error('Invalid tensor shape.');
+        }
         this._dimensions = dimensions;
     }
 

test/models.json

@@ -1834,7 +1834,7 @@
   {
     "type":   "darknet",
     "target": "alexnet.cfg,alexnet.weights",
-    "source": "https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/alexnet.cfg,https://pjreddie.com/media/files/alexnet.weights",
+    "source": "https://github.com/pjreddie/darknet/blob/master/cfg/alexnet.cfg?raw=true,https://pjreddie.com/media/files/alexnet.weights",
     "format": "Darknet",
     "link":   "https://pjreddie.com/darknet/imagenet"
   },
@@ -1855,7 +1855,7 @@
   {
     "type":   "darknet",
     "target": "darknet53_448.cfg,darknet53_448.weights",
-    "source": "https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/darknet53_448.cfg,https://pjreddie.com/media/files/darknet53_448.weights",
+    "source": "https://github.com/pjreddie/darknet/blob/master/cfg/darknet53_448.cfg?raw=true,https://pjreddie.com/media/files/darknet53_448.weights",
     "format": "Darknet",
     "link":   "https://pjreddie.com/darknet/imagenet"
   },
@@ -1866,6 +1866,13 @@
     "format": "Darknet",
     "link":   "https://github.com/lutzroeder/netron/issues/381"
   },
+  {
+    "type":   "darknet",
+    "target": "lstm.train.cfg",
+    "source": "https://github.com/AlexeyAB/darknet/blob/master/cfg/lstm.train.cfg?raw=true",
+    "format": "Darknet",
+    "link":   "https://github.com/AlexeyAB/darknet"
+  },
   {
     "type":   "darknet",
     "target": "mixnet_m_gpu.cfg",
@@ -1873,24 +1880,38 @@
     "format": "Darknet",
     "link":   "https://github.com/lutzroeder/netron/issues/381"
   },
+  {
+    "type":   "darknet",
+    "target": "msr_50.cfg",
+    "source": "https://github.com/AlexeyAB/darknet/blob/master/cfg/msr_50.cfg?raw=true",
+    "format": "Darknet",
+    "link":   "https://github.com/AlexeyAB/darknet"
+  },
   {
     "type":   "darknet",
     "target": "go.cfg,go.weights",
-    "source": "https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/go.cfg,https://pjreddie.com/media/files/go.weights",
+    "source": "https://github.com/pjreddie/darknet/blob/master/cfg/go.cfg?raw=true,https://pjreddie.com/media/files/go.weights",
     "format": "Darknet",
     "link":   "https://pjreddie.com/darknet/darkgo-go-in-darknet"
   },
+  {
+    "type":   "darknet",
+    "target": "gru.cfg",
+    "source": "https://github.com/pjreddie/darknet/blob/master/cfg/gru.cfg?raw=true",
+    "format": "Darknet",
+    "link":   "https://pjreddie.com/darknet"
+  },
   {
     "type":   "darknet",
     "target": "jnet-conv.cfg,jnet-conv.weights",
-    "source": "https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/jnet-conv.cfg,https://pjreddie.com/media/files/jnet-conv.weights",
+    "source": "https://github.com/pjreddie/darknet/blob/master/cfg/jnet-conv.cfg?raw=true,https://pjreddie.com/media/files/jnet-conv.weights",
     "format": "Darknet",
     "link":   "https://pjreddie.com/darknet/nightmare"
   },
   {
     "type":   "darknet",
     "target": "grrm.cfg,grrm.weights",
-    "source": "https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/rnn.cfg,https://pjreddie.com/media/files/grrm.weights",
+    "source": "https://github.com/pjreddie/darknet/blob/master/cfg/rnn.cfg?raw=true,https://pjreddie.com/media/files/grrm.weights",
     "format": "Darknet",
     "link":   "https://pjreddie.com/darknet/rnns-in-darknet"
   },
@@ -1904,24 +1925,66 @@
   {
     "type":   "darknet",
     "target": "resnet18.cfg,resnet18.weights",
-    "source": "https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/resnet18.cfg,https://pjreddie.com/media/files/resnet18.weights",
+    "source": "https://github.com/pjreddie/darknet/blob/master/cfg/resnet18.cfg?raw=true,https://pjreddie.com/media/files/resnet18.weights",
     "format": "Darknet",
     "link":   "https://pjreddie.com/darknet/imagenet"
   },
   {
     "type":   "darknet",
     "target": "resnet152.cfg,resnet152.weights",
-    "source": "https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/resnet152.cfg,https://pjreddie.com/media/files/resnet152.weights",
+    "source": "https://github.com/pjreddie/darknet/blob/master/cfg/resnet152.cfg?raw=true,https://pjreddie.com/media/files/resnet152.weights",
     "format": "Darknet",
     "link":   "https://pjreddie.com/darknet/imagenet"
   },
+  {
+    "type":   "darknet",
+    "target": "tiny.cfg,tiny.weights",
+    "source": "https://github.com/pjreddie/darknet/blob/master/cfg/tiny.cfg?raw=true,https://pjreddie.com/media/files/tiny.weights",
+    "format": "Darknet",
+    "link":   "https://pjreddie.com/darknet/tiny-darknet"
+  },
+  {
+    "type":   "darknet",
+    "target": "yolo-coco.cfg",
+    "source": "https://github.com/AlexeyAB/darknet/blob/master/cfg/yolov1/yolo-coco.cfg?raw=true",
+    "format": "Darknet",
+    "link":   "https://github.com/AlexeyAB/darknet"
+  },
+  {
+    "type":   "darknet",
+    "target": "yolo-voc.cfg",
+    "source": "https://github.com/AlexeyAB/darknet/blob/master/cfg/yolo-voc.cfg?raw=true",
+    "format": "Darknet",
+    "link":   "https://github.com/AlexeyAB/darknet"
+  },
+  {
+    "type":   "darknet",
+    "target": "yolo.cfg",
+    "source": "https://github.com/AlexeyAB/darknet/blob/master/cfg/yolo.cfg?raw=true",
+    "format": "Darknet",
+    "link":   "https://github.com/AlexeyAB/darknet"
+  },
+  {
+    "type":   "darknet",
+    "target": "yolo9000.cfg",
+    "source": "https://github.com/AlexeyAB/darknet/blob/master/cfg/yolo9000.cfg?raw=true",
+    "format": "Darknet",
+    "link":   "https://github.com/AlexeyAB/darknet"
+  },
   {
     "type":   "darknet",
     "target": "yolov3-spp.cfg,yolov3-spp.weights",
-    "source": "https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/yolov3-spp.cfg,https://pjreddie.com/media/files/yolov3-spp.weights",
+    "source": "https://github.com/pjreddie/darknet/blob/master/cfg/yolov3-spp.cfg?raw=true,https://pjreddie.com/media/files/yolov3-spp.weights",
     "format": "Darknet",
     "link":   "https://pjreddie.com/darknet/yolo"
   },
+  {
+    "type":   "darknet",
+    "target": "yolov3-tiny_occlusion_track.cfg",
+    "source": "https://github.com/AlexeyAB/darknet/blob/master/cfg/yolov3-tiny_occlusion_track.cfg?raw=true",
+    "format": "Darknet",
+    "link":   "https://github.com/AlexeyAB/darknet"
+  },
   {
     "type":   "dl4j",
     "target": "darknet19_dl4j_inference.v2.zip",

tools/darknet

@@ -14,7 +14,7 @@ clean() {
 
 sync() {
     bold "darknet sync"
-    [ -d "./third_party/src/darknet" ] || git clone --quiet https://github.com/pjreddie/darknet.git "./third_party/src/darknet"
+    [ -d "./third_party/src/darknet" ] || git clone --quiet https://github.com/AlexeyAB/darknet.git "./third_party/src/darknet"
     pushd "./third_party/src/darknet" > /dev/null
     git pull --quiet --prune
     popd > /dev/null